Periodic Reporting for period 2 - NeMeCo (Near Memory Computing)
Período documentado: 2018-04-01 hasta 2020-09-30
The key scientific objective of the NeMeCo project is to enable the design of a power-efficient High Performance Computing systems for big data type of applications based on Near-Memory Computing (NMC). Big data applications are often memory bound (i.e. memory access time dominates the other computation time) and/or require a huge memory bandwidth. Near-memory processing is one of the few real solutions to address the current scaling issues in HPC (High Performance Computing) systems in order to realize Exascale computers that are needed for near-future big data workloads. However, NMC is still in its infancy. Before it can be established as an essential component of HPC systems and be exploited for accelerating Big-data workloads, multiple challenges have to be addressed.
Besides the design of the NMC device itself, this includes 1) its integration into the overall computer system architecture, 2) investigate how multiple NMC devices can work together to scale to larger data volumes, and 3) how such a hybrid system can be effectively programmed to maximize performance and minimize power consumption at the system level. Key objectives for the 3 ESRs are to prepare and publish journal papers about:
1.New NMC architecture(s);
2.Adequate programming models for NMC;
3.New compiler technology for NMC;
4.New compile- and run-time optimization techniques for NMC;
5.Application of NMC to real-world problems.
The consortium is composed by two beneficiaries Eindhoven University of Technology (TU/e), IBM Research GmbH Zurich Research Laboratory (IBM) and three associative partner organisations: Dutch institute for radio astronomy (ASTRON), Swiss Federal Institute of Technology in Zurich (ETH-Z) and Dresden university of technology (TUDresden).
Besides the design of the NMC device itself, this includes 1) its integration into the overall computer system architecture, 2) investigate how multiple NMC devices can work together to scale to larger data volumes, and 3) how such a hybrid system can be effectively programmed to maximize performance and minimize power consumption at the system level. Key objectives for the 3 ESRs are to prepare and publish journal papers about:
1.New NMC architecture(s);
2.Adequate programming models for NMC;
3.New compiler technology for NMC;
4.New compile- and run-time optimization techniques for NMC;
5.Application of NMC to real-world problems.
The consortium is composed by two beneficiaries Eindhoven University of Technology (TU/e), IBM Research GmbH Zurich Research Laboratory (IBM) and three associative partner organisations: Dutch institute for radio astronomy (ASTRON), Swiss Federal Institute of Technology in Zurich (ETH-Z) and Dresden university of technology (TUDresden).
Many big data applications and memory bound applications will benefit from near-memory computing. The performance and power consumption of these kind of applications is dominated by the data transfer cost in traditional systems. Data transfer cost can be significantly reduced by moving the computation closer to the memory. To demonstrate the advantages of near-memory computing, the following big data application are demonstrated on the NeMeCo platform:
• Radio astronomy (W-Projection gridder and Correlator in the Square Kilometre array pipeline and De-dispersion algorithm in the Pulsar pipeline);
• Graph500 and Ligra (Graph Processing and Databases);
• Convolutional Neural Networks (Deep Learning);
• Nussinov’s Algorithm (Genomic Sequencing).
Cosmo weather model (weather forcasting)
NeMeCo research is driven by the current weak HPC performance for memory bound big data applications. The following results have been obtained by the ESRs in their individual projects:
ESR1: To characterize big data applications and to design runtime optimizations for selected applications to be executed on the NeMeCo platform, all kind of big data applications have been analysed with enhanced application modelling tools. Big data applications have been characterized in being “memory bound” or “not memory bound”. Next to this, the required memory bandwidth is determined for the computation of these applications on different styles of near memory computing systems.
ESR2: To design new compiler flows for near memory computing with dedicated optimizations and parallelizations for NeMeCo devices, literature and the software code of selected big data applications has been analysed in detail for the development of the NeMeCo compiler framework and tool flow. ESR2 analysed the selected big data applications from compiler point of view. Based on this analysis a new high-level optimization framework to improve the compilation process for these systems was presented.
ESR3: To design the NeMeCo architecture and accelerator the whole design space for NeMeCo architectures has been explored. NMC architectures have been explored by literature study, modelling, simulation and evaluation of prototype NeMeCo systems. A paper on classifying NMC systems has been written and published. ESR3 analysed the selected big data applications from an architectural point of view.
• Radio astronomy (W-Projection gridder and Correlator in the Square Kilometre array pipeline and De-dispersion algorithm in the Pulsar pipeline);
• Graph500 and Ligra (Graph Processing and Databases);
• Convolutional Neural Networks (Deep Learning);
• Nussinov’s Algorithm (Genomic Sequencing).
Cosmo weather model (weather forcasting)
NeMeCo research is driven by the current weak HPC performance for memory bound big data applications. The following results have been obtained by the ESRs in their individual projects:
ESR1: To characterize big data applications and to design runtime optimizations for selected applications to be executed on the NeMeCo platform, all kind of big data applications have been analysed with enhanced application modelling tools. Big data applications have been characterized in being “memory bound” or “not memory bound”. Next to this, the required memory bandwidth is determined for the computation of these applications on different styles of near memory computing systems.
ESR2: To design new compiler flows for near memory computing with dedicated optimizations and parallelizations for NeMeCo devices, literature and the software code of selected big data applications has been analysed in detail for the development of the NeMeCo compiler framework and tool flow. ESR2 analysed the selected big data applications from compiler point of view. Based on this analysis a new high-level optimization framework to improve the compilation process for these systems was presented.
ESR3: To design the NeMeCo architecture and accelerator the whole design space for NeMeCo architectures has been explored. NMC architectures have been explored by literature study, modelling, simulation and evaluation of prototype NeMeCo systems. A paper on classifying NMC systems has been written and published. ESR3 analysed the selected big data applications from an architectural point of view.
The key innovation is the development of runtime optimization, programming models, compilers and accelerators for near memory computing for a NeMeCo demonstrator.
The NeMeCo activities have the following impacts: human resource development for a next generation of NeMeCo computers, employability and career opportunities for the ESRs, multi-disciplinary research training at the doctoral level, stronger industry-academia connections, aligned R&D agendas, including training on the job, entrepreneurship and start-ups, and efficient design trajectories for High performance Computers.
A total of 13 scientific publications have been accepted.
The NeMeCo activities have the following impacts: human resource development for a next generation of NeMeCo computers, employability and career opportunities for the ESRs, multi-disciplinary research training at the doctoral level, stronger industry-academia connections, aligned R&D agendas, including training on the job, entrepreneurship and start-ups, and efficient design trajectories for High performance Computers.
A total of 13 scientific publications have been accepted.